NAACL 2024 论文列表

Understanding when two pieces of text convey the same information is a goal touching many subproblems in NLP, including textual entailment and fact-checking. This problem becomes more complex when those two pieces of text are in different languages. Here, we introduce X-PARADE (Cross-lingual Paragraph-level Analysis of Divergences and Entailments), the first cross-lingual dataset of paragraph-level information divergences. Annotators label a paragraph in a target language at the span level and evaluate it with respect to a corresponding paragraph in a source language, indicating whether a given piece of information is the same, new, or new but can be inferred. This last notion establishes a link with cross-language NLI. Aligned paragraphs are sourced from Wikipedia pages in different languages, reflecting real information divergences observed in the wild. Armed with our dataset, we investigate a diverse set of approaches for this problem, including classic token alignment from machine translation, textual entailment methods that localize their decisions, and prompting LLMs. Our results show that these methods vary in their capability to handle inferable information, but they all fall short of human performance.

The impressive development of large language models (LLMs) is expanding into the realm of large multimodal models (LMMs), which incorporate multiple types of data beyond text. However, the nature of multimodal models leads to significant expenses in the creation of training data. Furthermore, constructing multilingual data for LMMs presents its own set of challenges due to language diversity and complexity. Therefore, in this study, we propose two cost-effective methods to solve this problem: (1) vocabulary expansion and pretraining of multilingual LLM for specific languages, and (2) automatic and elaborate construction of multimodal datasets using GPT4-V. Based on these methods, we constructed a 91K English-Korean-Chinese multilingual, multimodal training dataset. Additionally, we developed a bilingual multimodal model that exhibits excellent performance in both Korean and English, surpassing existing approaches.

Natural Language Generation (NLG) typically involves evaluating the generated text in various aspects (e.g., consistency and naturalness) to obtain a comprehensive assessment. However, multi-aspect evaluation remains challenging as it may require the evaluator to generalize to any given evaluation aspect even if it’s absent during training. In this paper, we introduce X-Eval, a two-stage instruction tuning framework to evaluate text in both seen and unseen aspects customized by end users. X-Eval consists of two learning stages: the vanilla instruction tuning stage that improves the model’s ability to follow evaluation instructions, and an enhanced instruction tuning stage that exploits the connections between fine-grained evaluation aspects to better assess text quality. To support the training of X-Eval, we collect AspectInstruct, the first instruction tuning dataset tailored for multi-aspect NLG evaluation spanning 27 diverse evaluation aspects with 65 tasks. To enhance task diversity, we devise an augmentation strategy that converts human rating annotations into diverse forms of NLG evaluation tasks, including scoring, comparison, ranking, and Boolean question answering. Extensive experiments across three essential categories of NLG tasks: dialogue generation, summarization, and data-to-text coupled with 21 aspects in meta-evaluation, demonstrate that X-Eval enables even a lightweight language model to achieve a comparable if not higher correlation with human judgments compared to the state-of-the-art NLG evaluators like GPT-4.

Most existing retrieval-augmented language models (LMs) assume a naive dichotomy within a retrieved document set: query-relevance and irrelevance. Our work investigates a more challenging scenario in which even the “relevant” documents may contain misleading or incorrect information, causing conflict among the retrieved documents and thereby negatively influencing model decisions as noise. We observe that existing LMs are highly brittle to the presence of conflicting information in both the fine-tuning and in-context few-shot learning scenarios. We propose approaches for handling knowledge conflicts among retrieved documents by explicitly fine-tuning a discriminator or prompting GPT-3.5 to elicit its discriminative capability. Our empirical results on open-domain QA show that these approaches significantly enhance model robustness. We also provide our findings on incorporating the fine-tuned discriminator’s decision into the in-context learning process, proposing a way to exploit the benefits of two disparate learning schemes. Alongside our findings, we provide MacNoise, a machine-generated, conflict-induced dataset to further encourage research in this direction.

*Do larger and more performant models resolve NLP’s longstanding robustness issues?* We investigate this question using over 20 models of different sizes spanning different architectural choices and pretraining objectives. We conduct evaluations using (a) out-of-domain and challenge test sets, (b) behavioral testing with CheckLists, (c) contrast sets, and (d) adversarial inputs. Our analysis reveals that not all out-of-domain tests provide insight into robustness. Evaluating with CheckLists and contrast sets shows significant gaps in model performance; merely scaling models does not make them adequately robust. Finally, we point out that current approaches for adversarial evaluations of models are themselves problematic: they can be easily thwarted, and in their current forms, do not represent a sufficiently deep probe of model robustness. We conclude that not only is the question of robustness in NLP as yet unresolved, but even some of the approaches to measure robustness need to be reassessed.

When connecting objects and their language referents in an embodied 3D environment, it is important to note that: (1) an object can be better characterized by leveraging comparative information between itself and other objects, and (2) an object’s appearance can vary with camera position. As such, we present the Multi-view Approach to Grounding in Context (MAGiC) model, which selects an object referent based on language that distinguishes between two similar objects. By pragmatically reasoning over both objects and across multiple views of those objects, MAGiC improves over the state-of-the-art model on the SNARE object reference task with a relative error reduction of 12.9% (representing an absolute improvement of 2.7%). Ablation studies show that reasoning jointly over object referent candidates and multiple views of each object both contribute to improved accuracy. Code: https://github.com/rcorona/magic_snare/

Our research integrates graph data with Large Language Models (LLMs), which, despite their advancements in various fields using large text corpora, face limitations in encoding entire graphs due to context size constraints. This paper introduces a new approach to encoding a graph with diverse modalities, such as text, image, and motif, coupled with prompts to approximate a graph’s global connectivity, thereby enhancing LLMs’ efficiency in processing complex graph structures. The study also presents GraphTMI, a novel benchmark for evaluating LLMs in graph structure analysis, focusing on homophily, motif presence, and graph difficulty. Key findings indicate that the image modality, especially with vision-language models like GPT-4V, is superior to text in balancing token limits and preserving essential information and comes close to prior graph neural net (GNN) encoders. Furthermore, the research assesses how various factors affect the performance of each encoding modality and outlines the existing challenges and potential future developments for LLMs in graph understanding and reasoning tasks. Our code and data are publicly available on our project page - https://minnesotanlp.github.io/GraphLLM/

We train models to answer the question, Where are you from? and show how such models can be repurposed for language identification (LID). To our knowledge, this paper is the first to introduce data sources, methods and models to tackle the task of geolocation of speech at a global scale, and the first to explore using geolocation as a proxy-task for LID. Specifically, we explore whether radio broadcasts with known origin can be used to train regression and classification-based models for geolocating speech. We build models on top of self-supervised pretrained models, using attention pooling to qualitatively verify that the model geolocates the speech itself, and not other channel artifacts.The best geolocation models localize speaker origin to around 650km. We confirm the value of speech geolocation as a proxy task by using speech geolocation models for zero-shot LID. Finally, we show that fine-tuning geolocation models for LID outperforms fine-tuning pretrained Wav2Vec2.0 models, and achieves state-of-the-art performance on the FLEURS benchmark.

Recent studies introduced effective compression techniques for Large Language Models (LLMs) via post-training quantization or low-bit weight representation. Although quantized weights offer storage efficiency and allow for faster inference, existing works have indicated that quantization might compromise performance and exacerbate biases in LLMs.This study investigates the confidence and calibration of quantized models, considering factors such as language model type and scale as contributors to quantization loss.Firstly, we reveal that quantization with GPTQ to 4-bit results in a decrease in confidence regarding true labels, with varying impacts observed among different language models. Secondly, we observe fluctuations in the impact on confidence across different scales. Finally, we propose an explanation for quantization loss based on confidence levels, indicating that quantization disproportionately affects samples where the full model exhibited low confidence levels in the first place.We make our code and quantized models publicly available.

Deployed dialogue agents have the potential to integrate human feedback to continuously improve themselves. However, humans may not always provide explicit signals when the chatbot makes mistakes during interactions. In this work, we propose Juicer, a framework to make use of both binary and free-form textual human feedback. It works by: (i) extending sparse binary feedback by training a satisfaction classifier to label the unlabeled data; and (ii) training a reply corrector to map the bad replies to good ones. We find that augmenting training with model-corrected replies improves the final dialogue model, and we can further improve performance by using both positive and negative replies through the recently proposed Director model.

Recent studies suggest that self-reflective prompting can significantly enhance the reasoning capabilities of Large Language Models (LLMs). However, the use of external feedback as a stop criterion raises doubts about the true extent of LLMs’ ability to emulate human-like self-reflection. In this paper, we set out to clarify these capabilities under a more stringent evaluation setting in which we disallow any kind of external feedback. Our findings under this setting show a split: while self-reflection enhances performance in TruthfulQA, it adversely affects results in HotpotQA.We conduct follow-up analyses to clarify the contributing factors in these patterns, and find that the influence of self-reflection is impacted both by reliability of accuracy in models’ initial responses, and by overall question difficulty: specifically, self-reflection shows the most benefit when models are less likely to be correct initially, and when overall question difficulty is higher. We also find that self-reflection reduces tendency toward majority voting. Based on our findings, we propose guidelines for decisions on when to implement self-reflection. We release the codebase for reproducing our experiments at https://github.com/yanhong-lbh/LLM-SelfReflection-Eval.

Multilingual machine translation (MMT), trained on a mixture of parallel and monolingual data, is key for improving translation in low-resource language pairs. However, the literature offers conflicting results on the performance of different methods of including monolingual data. To resolve this, we examine how denoising autoencoding (DAE) and backtranslation (BT) impact MMT under different data conditions and model scales. Unlike prior studies, we use a realistic dataset of 100 translation directions and consider many domain combinations of monolingual and test data. We find that monolingual data generally helps MMT, but models are surprisingly brittle to domain mismatches, especially at smaller model scales. BT is beneficial when the parallel, monolingual, and test data sources are similar but can be detrimental otherwise, while DAE is less effective than previously reported. Next, we analyze the impact of scale (from 90M to 1.6B parameters) and find it is important for both methods, particularly DAE. As scale increases, DAE transitions from underperforming the parallel-only baseline at 90M to converging with BT performance at 1.6B, and even surpassing it in low-resource. These results offer new insights into how to best use monolingual data in MMT.

Eliciting feedback from end users of NLP models can be beneficial for improving models. However, how should we present model responses to users so they are most amenable to be corrected from user feedback? Further, what properties do users value to understand and trust responses? We answer these questions by analyzing the effect of rationales (or explanations) generated by QA models to support their answers. We specifically consider decomposed QA models that first extract an intermediate rationale based on a context and a question and then use solely this rationale to answer the question. A rationale outlines the approach followed by the model to answer the question. Our work considers various formats of these rationales that vary according to well-defined properties of interest. We sample rationales from language models using few-shot prompting for two datasets, and then perform two user studies. First, we present users with incorrect answers and corresponding rationales in various formats and ask them to provide natural language feedback to revise the rationale. We then measure the effectiveness of this feedback in patching these rationales through in-context learning. The second study evaluates how well different rationale formats enable users to understand and trust model answers, when they are correct. We find that rationale formats significantly affect how easy it is (1) for users to give feedback for rationales, and (2) for models to subsequently execute this feedback. In addition, formats with attributions to the context and in-depth reasoning significantly enhance user-reported understanding and trust of model outputs.

Recent advancements in GPT-4V have displayed remarkable multi-modal capabilities in processing image inputs and following open-ended instructions. Despite these advancements, there is considerable scope for enhancing open-source multi-modal LLMs, especially in terms of multi-modal understanding accuracy and instruction-following proficiency. In this paper, we conduct a comprehensive study on training GPT4-style models. We introduce Lynx a multi-modal LLM developed through a series of controlled experiments comparing various model variants. This process allowed us to identify and implement an optimal training strategy tailored for multi-modal LLMs. In addition to our model development, we propose a plug-and-play technique designed to augment the instruction-following capabilities of multi-modal LLMs. We have validated the performance of Lynx on multiple benchmarks. Results demonstrate that Lynx not only achieves strong image understanding accuracy but also excels in instruction-following tasks, paving the path for ongoing enhancements in multi-modal LLMs.

This paper investigates the question of what makes math word problems (MWPs) in English challenging for large language models (LLMs). We conduct an in-depth analysis of the key linguistic and mathematical characteristics of MWPs. In addition, we train feature-based classifiers to better understand the impact of each feature on the overall difficulty of MWPs for prominent LLMs and investigate whether this helps predict how well LLMs fare against specific categories of MWPs.

We consider an unanswered question in the discourse processing community: why do relation classifiers trained on explicit examples (with connectives removed) perform poorly in real implicit scenarios? Prior work claimed this is due to linguistic dissimilarity between explicit and implicit examples but provided no empirical evidence. In this study, we show that one cause for such failure is a label shift after connectives are eliminated. Specifically, we find that the discourse relations expressed by some explicit instances will change when connectives disappear. Unlike previous work manually analyzing a few examples, we present empirical evidence at the corpus level to prove the existence of such shift. Then, we analyze why label shift occurs by considering factors such as the syntactic role played by connectives, ambiguity of connectives, and more. Finally, we investigate two strategies to mitigate the label shift: filtering out noisy data and joint learning with connectives. Experiments on PDTB 2.0, PDTB 3.0, and the GUM dataset demonstrate that classifiers trained with our strategies outperform strong baselines.

Vision-language (VL) models, pretrained on colossal image-text datasets, have attained broad VL competence that is difficult to evaluate. A common belief is that a small number of VL skills underlie the variety of VL tests. In this paper, we perform a large-scale transfer learning experiment aimed at discovering latent VL skills from data. We reveal interesting characteristics that have important implications for test suite design. First, generation tasks suffer from a length bias, suggesting benchmarks should balance tasks with varying output lengths. Second, we demonstrate that factor analysis successfully identifies reasonable yet surprising VL skill factors, suggesting benchmarks could leverage similar analyses for task selection.Finally, we present a new dataset, OLIVE1, which simulates user instructions in the wild and presents challenges dissimilar to all datasets we tested. Our findings contribute to the design of balanced and broad-coverage vision-language evaluation methods. 1https://github.com/jq-zh/olive-dataset

Knowledge Distillation (KD) is a predominant approach for BERT compression.Previous KD-based methods focus on designing extra alignment losses for the student model to mimic the behavior of the teacher model.These methods transfer the knowledge in an indirect way.In this paper, we propose a novel Weight-Inherited Distillation (WID), which directly transfers knowledge from the teacher.WID does not require any additional alignment loss and trains a compact student by inheriting the weights, showing a new perspective of knowledge distillation.Specifically, we design the row compactors and column compactors as mappings and then compress the weights via structural re-parameterization.Experimental results on the GLUE and SQuAD benchmarks show that WID outperforms previous state-of-the-art KD-based baselines.Further analysis indicates that WID can also learn the attention patterns from the teacher model without any alignment loss on attention distributions.The code is available at https://github.com/wutaiqiang/WID-NAACL2024.

This paper investigates using Large Language Models (LLMs) to automatically perform web software tasks using click, scroll, and text in- put operations. Previous approaches, such as reinforcement learning (RL) or imitation learning, are inefficient to train and task-specific. Our method uses filtered Document Object Model (DOM) elements as observations and performs tasks step-by-step, sequentially generating small programs based on the current observations. We use in-context learning, either benefiting from a single manually provided example, or an automatically generated example based on a successful zero-shot trial. We evaluate our proposed method on the MiniWob++ benchmark. With only one in-context example, our WebWISE method using gpt-3.5-turbo achieves similar or better performance than other methods that require many demonstrations or trials.